A novel approach of initializing region-based active contours in noisy images by means of higher order statistics and dissimilarity

In this paper we present two novel methods for initializing region-based active contours. First we use an analysis of higher order statistic attributes of the complete image and the regions of the initial contour candidates. Then we extend this approach by taking the dissimilarity between the initialized regions and the complete image into account. We compare our method with the standard initialization of active contours for different types and degree of noise in synthetic and real images. It is shown that our method outperforms state of the art initializations of active contours considering the total accuracy and Cohen's kappa.